AI Found Thousands of Software Flaws, but Very Few Have Been Fixed, Researcher Warns
A new analysis from VulnCheck vulnerability researcher Patrick Garrity questions how much real-world benefit Anthropic's Project Glasswing has delivered. The project gives security researchers access to Anthropic's advanced AI models to help find and fix software flaws. Seven months on, Garrity says maintainers are being "overwhelmed" by "underwhelming findings". His report, released on November 9, draws on data Anthropic has published itself.
According to Anthropic's disclosure dashboard dated November 2, 2026, researchers using Glasswing have found 37,966 potential vulnerabilities. Only 7,349 (19.4%) appear in Anthropic's public ledger. Garrity found 5,276 of those were validated as real, unique vulnerabilities, while 1,654 were duplicates and 419 were withdrawn. Of the validated flaws, 4,866 have been reported to maintainers and 193 are awaiting disclosure. Just 217 have been fixed, which is less than 0.5% of the total claimed. Since September, 4,613 findings were added to the ledger but only 15 were marked as fixed.
Garrity also questions Anthropic's claimed true positive rate of 94.2%, saying it raises questions about how duplicates and withdrawn findings are counted. One possible reason for the gap between findings and fixes is the pace at which people can triage vulnerability reports. This is only the opening of the source article, so the full report may contain further detail.